Thursday, May 31, 2012

Reverse Engineering the Birthday Paradox

Recently, I've been thinking about the Birthday Paradox (not to be confused with this similarly named problem). The first time was while reviewing the concept of birthday attacks for the final exam of the online cryptography class I'm currently taking. The second time was while reading this intriguing exploration of the processes at play when struggling to understand a math problem.

The usual way to think about the problem is to consider the probability $\overline{p}$ of the opposite event we're interested in: that there is no birthday"collision" at all within a group of $k$ people. This, we're said (without really explaining why), is easier to calculate, and it follows from it that we can retrieve the probability $p$ of our goal event (i.e. that there is at least one "collision") as $1 - \overline{p}$.

The above-mentioned essay discusses at length this idea and the underlying principles, but it also casually challenges the reader with an interesting side problem: find a way to calculate $p$ directly, that is, without relying on the probability of the opposite event. Being a programmer, I decided to tackle this problem not from the principles, but from the opposite direction, by trying to derive understanding from "procedural tinkering".

In the first version of this article, I had derived a way of screening the $k$-permutations out of the cartesian power set $N^k$. I thought this was the answer, but someone on the Cryptography Class discussion forum helped me understand that this was actually only a rearrangement of the indirect computation (i.e. $p = 1 - \overline{p}$). The correct way to compute $p$ directly should rather involve the sum of:
  • the chance of a collision occurring only on the second choice
  • the chance of a collision occurring only on the third choice
  • ...
  • the chance of a collision occurring only on the $k$-th choice
This seemed to make sense, but as I wanted to study this proposition in more detail on a simplified problem instance, I wrote this Python program:

from __future__ import division
from itertools import product

n = 10
k = 5

prev_colls = set()
p_coll = 0

def findPrevColl(p, j):
    for i in range(2, j):
        if p[:i] in prev_colls:
            return True
    return False

for j in range(2, k+1):
    n_colls = 0
    count = 0
    for p in product(range(n), repeat=j):
        if len(set(p)) < j and not findPrevColl(p, j):
            n_colls += 1
            prev_colls.add(p)
        count += 1
    print 'at k = %d, P_k = %f' % (j, n_colls / count)
    p_coll += n_colls / count
print 'sum(P_k) = %f' % p_coll

# verify result with the "indirect" formula
import operator
from numpy.testing import assert_approx_equal
assert_approx_equal(p_coll, 1 - reduce(operator.mul, range(n-k+1, n+1)) / n ** k)

# at k = 2, P_k = 0.100000
# at k = 3, P_k = 0.180000
# at k = 4, P_k = 0.216000
# at k = 5, P_k = 0.201600
# sum(P_k) = 0.697600

which works by enumerating, for every $k$, all the possible trials (of length $k$) and looking for a "new" collision with each one, "new" meaning that no subset (of length $< k$) of this particular trial has ever been causing a collision. The probabilities for each $k$ are then summed to yield the overall, direct probability of at least one collision. Note that this code relies on the itertools module's product function to generate the cartesian powers of $k$ (i.e. all possible trials of length $k$) and the set data structure for easy past and current collision detection. Once fully convinced that this was working, the obvious next step was to derive an analytic formulation for it. By studying some actual values from my program, I figured out that the probability for a collision occurring only on the $k$-th choice should be: \[ P_{only}(n, k) = \frac{\left(\frac{n! \cdot (k-1)}{(n-k+1)!}\right)}{n^k}\] meaning that the total, direct probability of at least one collision is the sum: \[ P_{any}(n, k) = \sum_k{\frac{\left(\frac{n! \cdot (k-1)}{(n-k+1)!}\right)}{n^k}}\] As this wasn't still fully satisfying, because it doesn't yield an intuitive understanding of what's happening, the same helpful person on the Cryptography forum offered an equivalent, but much better rewrite: \[ P_{only}(n, k) = \left(\frac{n}{n} \cdot \frac{n-1}{n} \cdot \cdot \cdot \frac{n-k+2}{n}\right) \cdot \left(\frac{k-1}{n}\right)\] which can be easily understood by imagining a bucket of $n$ elements: the chance of a collision happening exactly at choice $k$ is the probability that there was no collision with the first $k-1$ choices ($\frac{n}{n} \cdot \frac{n-1}{n} \cdot \cdot \cdot$, increasing as the bucket fills up) multiplied by the probability of a collision at $k$, which will happen $k-1$ times out of $n$, if we assume that the bucket is already filled with $k-1$ different values (by virtue of the no-previous-collision assumption).

Although the conclusion of my previous attempt was that it's often possible to derive mathematical understanding from "procedural tinkering" (i.e. from programming to math), I'm not so sure anymore, as this second version is definitely a counterexample.